LSTM Derin Öğrenme Yaklaşımı ile Covid-19  Pandemi Sürecinde Twitter Verilerinden Duygu  Analizi

Yılmaz, Mehmet; Orman, Zeynep

doi:10.26650/acin.947747

Cited by 6 publications

(9 citation statements)

References 1 publication

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…In the scope of KTS process [62], related video segments (i.e., video shots) needed for video summarization have been chosen based on change points that have been detected. This meta model was built with VGG19 [60] and Bi-LSTM [14][15][16] parts as told in the above sections of this study. After this kind of summarization training, the original full length video may come as an input when it is used to summarize with the use of shot level scores as average frames scores.…”

Section: Multimodal Deep Learning Methodologymentioning

confidence: 99%

“…Shel et al [17], proposed a model which is one of the examples for multimodal-based video prediction using deep learning. The idea behind their work is to create CNN [4] features from the frames of the video and give them to generated Long Short Term Memory (LSTM) [14][15][16] layers and to create MFCC features from the audio tracks of these video frames and give them to a generated different LSTM layers. Therefore, multimodal feature learning is applied into these two outputs and the classification process is done at the end.…”

Section: Related Workmentioning

confidence: 99%

“…By using both obtained importance scores and detected change points (i.e., temporal segmentation) within the scope of KTS process, thus related video segments needed for video summarization have been chosen. For long-term temporal dependencies between video frames and shots located far away each other, the above mentioned model's frame-relation learner part is used to build data for dynamic programming (i.e., Knapsack problem) that this part of the model is based on the bidirectional Long-Short Term Memory (Bi-LSTM) [14][15][16] and its structure was built on two Bi-LSTM layers which have 256 and 128 neural nodes, respectively. These layers work as a processor for revealing the forward hidden states and backward hidden states represented in a sequence of frames of video.…”

Section: Video Summarizationmentioning

confidence: 99%

“…The performance evaluation is based on the comparison between automatic summaries and ground truth summaries. In our study, an average performance score (i.e., F-metric) [3] was obtained as equals to 51% with the summarization tests on above mentioned TvSum and SumMe datasets with the integrated meta model's VGG19 [60], Bi-LSTM [14][15][16] and dynamic programming parts as told above. This performance score is the same as given in Zhou et al [61] study that they obtained this kind of average score in their study as equal to 50% (i.e., for SumMe as 42.1% and TVSum as 58.1%).…”

Section: Video Summarizationmentioning

confidence: 99%

“…In addition, the second contribution of our study is applying sound scene classification and recognition to reach a refined sentiment analysis from spoken content that related sentences are taken from the processed video clip using speech recognition, natural language processing (NLP) and ARM techniques. In the literature, there are several analysis methods for sentimental context that are pretrained and used with some published studies [13,14]. Everyone could choose any one of this kind of pre-trained model by respect to its performance, by its language type, by its emotional representation styles (negative/positive speech, angry/happy speech etc.)…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

An Approach for Audio-Visual Content Understanding of Video using Multimodal Deep Learning Methodology

BOZTEPE

KARAKAYA

Karasulu

et al. 2022

Sakarya University Journal of Computer and Information Sciences

View full text Add to dashboard Cite

This study contains an approach for recognizing the sound environment class from a video to understand the spoken content with its sentimental context via some sort of analysis that is achieved by the processing of audio-visual content using multimodal deep learning methodology. This approach begins with cutting the parts of a given video which the most action happened by using deep learning and this cutted parts get concanarated as a new video clip. With the help of a deep learning network model which was trained before for sound recognition, a sound prediction process takes place. The model was trained by using different sound clips of ten different categories to predict sound classes. These categories have been selected by where the action could have happened the most. Then, to strengthen the result of sound recognition if there is a speech in the new video, this speech has been taken. By using Natural Language Processing (NLP) and Named Entity Recognition (NER) this speech has been categorized according to if the word of a speech has connotation of any of the ten categories. Sentiment analysis and Apriori Algorithm from Association Rule Mining (ARM) processes are preceded by identifying the frequent categories in the concanarated video and helps us to define the relationship between the categories owned. According to the highest performance evaluation values from our experiments, the accuracy for sound environment recognition for a given video's processed scene is 70%, average Bilingual Evaluation Understudy (BLEU) score for speech to text with VOSK speech recognition toolkit's English language model is 90% on average and for Turkish language model is 81% on average. Discussion and conclusion based on scientific findings are included in our study.

show abstract

Section: Multimodal Deep Learning Methodologymentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: Video Summarizationmentioning

confidence: 99%

Section: Video Summarizationmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

An Approach for Audio-Visual Content Understanding of Video using Multimodal Deep Learning Methodology

BOZTEPE

KARAKAYA

Karasulu

et al. 2022

Sakarya University Journal of Computer and Information Sciences

View full text Add to dashboard Cite

show abstract

Analysis of the Covid-19 Process in Terms of Health Managers

Palavar¹,

Çelik²,

Albayrak³

et al. 2022

2022 2nd International Conference on Computing and Machine Intelligence (ICMI)

View full text Add to dashboard Cite

Naive Bayes Sınıflandırıcısı Kullanılarak YouTube Verileri Üzerinden Çok Dilli Duygu Analizi

Şahınaslan

DALYAN

Şahinaslan

2022

Bilişim Teknolojileri Dergisi

View full text Add to dashboard Cite

Bu çalışma, Naive Bayes sınıflandırma algoritması kullanılarak sosyal medya verileri üzerinden dünya dillerinde yapılan yorum ve paylaşımların anlaşılmasına katkıda bulunmayı amaçlamaktadır. Duygu analizi veya fikir madenciliği, doğal dil işleme konularının bir alt bölümünde yer alır ve sosyal medya üzerinde yer alan verilerin anlamlandırabilmesi sağlanır. Yapay zekâ ve makine öğrenmesi gibi konularla birlikte günümüzde son derece popüler bir alandır. Küresel bir köye dönüşen dünyada, insanlar sosyal medya araçları üzerinden görüş, düşünce ve o konu hakkında memnuniyet durumları gibi pek çok veriyi paylaşmaktadır. Anlamlandırılmayı bekleyen bu veriler üzerinden insanların duygu ve düşünceleri ortaya çıkartabilmek mümkündür. Son dönemlerde Türkçe dilinde bir takım duygu analizi çalışmalarının yapıldığı görülmektedir. Ancak dünyadaki farklı dil ve lehçelerde yapılan yorumlar üzerinden çok dilli duygu analiz çalışması sınırlıdır. Bu çalışmada, Türkçe, İngilizce, Almanca, Fransızca, Arapça, Rusça ve Korece gibi farklı dünya dillerinden YouTube yorum verileri kullanılmıştır. PHP ile elde edilen bu ham veriler üzerinden anlamsız veriler temizlenerek belirli bir kalitede veri seti elde edilmiştir. Bu yorum verilerini İngilizce'ye çevirmek için Google çeviri aracı API'sı kullanılmıştır. Duygular, PHP Niiknow kütüphanesi üzerinden Naive Bayes algoritması kullanılarak pozitif, negatif ve nötr olarak sınıflandırılır. WEKA ile yapılan veri analizinde doğru sınıflandırma oranı %65,56 olarak bulunmuştur. Bu çalışma, 15.082 veri seti ve Google Çeviri aracı tarafından desteklenen 108 dil sayısı ile sınırlıdır.

show abstract

LSTM Derin Öğrenme Yaklaşımı ile Covid-19 Pandemi Sürecinde Twitter Verilerinden Duygu Analizi

Cited by 6 publications

References 1 publication

An Approach for Audio-Visual Content Understanding of Video using Multimodal Deep Learning Methodology

An Approach for Audio-Visual Content Understanding of Video using Multimodal Deep Learning Methodology

Analysis of the Covid-19 Process in Terms of Health Managers

Naive Bayes Sınıflandırıcısı Kullanılarak YouTube Verileri Üzerinden Çok Dilli Duygu Analizi

Contact Info

Product

Resources

About